+ All Categories
Home > Documents > Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018...

Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018...

Date post: 20-May-2020
Category:
Upload: others
View: 3 times
Download: 0 times
Share this document with a friend
33
Housing Price Prediction An Nguyen March 20, 2018 Abstract This paper explores the question of how house prices in five different counties are affected by housing characteristics (both internally, such as number of bathrooms, bed- rooms, etc. and externally, such as public schools’ scores or the walkability score of the neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin, three prominent housing websites, this paper utilizes both the hedonic pricing model (Lin- ear Regression) and various machine learning algorithms, such as Random Forest (RF) and Support Vector Regression (SVR), to predict house prices. The models’ prediction scores, as well as the ratio of overestimated houses to underestimated houses are com- pared against Zillow’s price estimation scores and ratio. Results show that SVR gives a better price prediction score than the Zillow’s baseline on the same dataset of Hunt County (TX) and RF gives close or the same prediction scores to the baseline on three other counties. Moreover, this paper’s models reduce the overestimated to underes- timated house ratio of 3:2 from Zillow’s estimation to a ratio of 1:1. This paper also identifies the four most important attributes in housing price prediction across the counties as assessment, comparable houses’ sold price, listed price and number of bathrooms. i
Transcript
Page 1: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

Housing Price Prediction

An Nguyen

March 20, 2018

Abstract

This paper explores the question of how house prices in five different counties are

affected by housing characteristics (both internally, such as number of bathrooms, bed-rooms, etc. and externally, such as public schools’ scores or the walkability score of theneighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin, three

prominent housing websites, this paper utilizes both the hedonic pricing model (Lin-

ear Regression) and various machine learning algorithms, such as Random Forest (RF)

and Support Vector Regression (SVR), to predict house prices. The models’ prediction

scores, as well as the ratio of overestimated houses to underestimated houses are com-

pared against Zillow’s price estimation scores and ratio. Results show that SVR gives

a better price prediction score than the Zillow’s baseline on the same dataset of Hunt

County (TX) and RF gives close or the same prediction scores to the baseline on three

other counties. Moreover, this paper’s models reduce the overestimated to underes-

timated house ratio of 3:2 from Zillow’s estimation to a ratio of 1:1. This paper also

identifies the four most important attributes in housing price prediction across the

counties as assessment, comparable houses’ sold price, listed price and number of bathrooms.

i

Page 2: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

ACKNOWLEDGEMENTS:

I would like to thank my thesis advisors - Prof. Chris Fernandes, Nick Webb, and HarlanHolt - for their advice and inputs on this project. Many thanks to my friends and familyas well, who spent countless hours to listen and provide feedbacks.

ii

Page 3: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

Contents

1 Introduction 1

2 Literature Review 3

3 Methods 5

3.1 Hedonic Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.2 Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3.2.1 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3.2.2 Support Vector Regression . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Data 10

4.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4.2 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4.3 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

5 Results 14

5.1 Prediction Scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

5.2 Overestimation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5.3 Attributes’ Effects On Sold Price . . . . . . . . . . . . . . . . . . . . . . . . . 19

6 Conclusion 20

7 Appendix 21

7.1 Surrogate Split . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

7.2 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

iii

Page 4: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

1 Introduction

According to the US Census Bureau, 560,000 houses were sold in the United States in

2016 [11]. In addition, 65% of all American families owned houses in 2016 [12]. For the

Americans who sold and bought these houses, a good housing price prediction would

better prepare them for what to expect before they make one of the most important finan-

cial decisions in their lives. A recent report from the Zillow Group, a popular housing

database website, indicates that house sellers and buyers are increasingly turning to on-

line research in order to estimate house price before contacting real estate agents [4]. Re-

searching how much the house you are interested in is worth on your own can be difficult

for multiple reasons. One particular reason is that there many factors that influence the

potential price of a house, making it more complicated for an individual to decide how

much a house is worth on their own without external help. This can lead to people mak-

ing poorly informed decisions about whether to buy or sell their houses and which prices

are reasonable. Because houses are long term investments, it is imperative that people

make their decisions with the most accurate information possible. Therefore, housing

websites such as Zillow, Trulia and Redfin 1, exist to provide estimations of housing val-

uations based on the houses’ characteristics, at no cost.

However, the estimations provided by these housing websites are not always accurate.

For example, Zillow states that their housing price prediction algorithm, called “Zesti-

mate”, only estimates 54.4% of houses within the 5% of their actual sale prices [22]. For

Trulia, only 48.2% of houses have Trulia-estimated prices to be within the 5% range of

their actual sold prices [20]. Therefore, the first question of this project is whether I can

outperform Zestimate’s prediction score or come close to it. In this project, I define the

prediction score as the percentage of houses whose estimated prices fall within the 5%

range of their actual sold prices. Using this project’s datasets and Zestimates as the pre-

1Zillow: https://www.zillow.com; Trulia: https://www.trulia.com; Redfin: https://www.redfin.com.

1

Page 5: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

dictions, I compute Zillow’s prediction scores and use them as the baselines to see how

well my own models perform. I chose Zillow’s estimator as a benchmark instead of its

competitors’ because Zillow is widely regarded as the most popular housing website due

to its large databases of 110 million houses and their 11 years of expertise in pricing es-

timations. According to Hitwise, a consumer analytics company, Zillow’s market share,

based on online visits to the site, is 27.2% in 2016, while the numbers for Trulia and Redfin

are 9.4% and 3.7%, respectively.

Zillow tends to overestimate their listed properties, meaning the Zestimates are higher

than the actual sold prices of the houses. In the dataset of 1,457 sold houses I collected,

the ratio of overestimated houses to underestimated houses is 3 to 2. Hollas, Rutherford

and Thomson (2010) studies Zillow’s estimations of single family houses and finds that

80% of their housing sample gathered from Zillow are overpriced by Zestimate [8]. For

a house seller who prices his house based on Zillow’s suggestion, he/she is likely to list

his/her house for more than what it is worth. According to a Zillow research in 2016, if a

house is priced above its true market valuation, it tends to stay on the market five times

longer compared to a house that is well-priced, suggesting a string penalty for overpric-

ing houses [19]. Moreover, the same research suggests that houses that have been on the

market for two months can lose 5% of its original listed price. Asabere and Huffman

(1993) also supports the theory of a reversed correlation between a house’s time on the

market and its final sold price [1]. Therefore, the second question of this project is whether

my models can get rid of this overestimation problem.

The final question of this project is what the most important factors affecting housing

prices are. In order to answer the three questions listed above, this project proposes using

both the hedonic pricing model and various machine learning algorithms.

2

Page 6: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

2 Literature Review

Sirmans, Macpherson and Zietz (2005) provides a study of 125 papers that use hedonic

pricing model to estimate house prices in the past decade [16]. The paper provides a list

of 20 attributes that are frequently used to specify hedonic pricing models. This dataset

contains 12 attributes on this list. Moreover, Sirmans, Macpherson, and Zietz (2005) also

discusses the effects of some variables on housing price. For example, number of bath-

rooms is usually positively correlated to the final sale price. Out of 40 times appearing in

housing price studies, this attribute has a positive effect 34 times and is statistically signifi-

cant 35 times. On average, keeping other variables unchanged, an increase of 1 bathroom

leads to 10% to 12% increase in the property’s value. Similarly, my paper shows that,

based on the dataset of sold houses in five counties, the number of bathroom has a statis-

tically significant and positive effect on sold price. On average, an increase of 1 bathroom

could increase a house’s price by $15,787.

Cebula (2009) conducts a study on the housing prices in the City of Savannah, Georgia

using the hedonic pricing model [3]. The paper’s data contains 2,888 single-family houses

for the period between 2000 and 2005. Cebula (2009) shows that the log price of houses

is positively and significantly correlated with the number of bathrooms, bedrooms, fire-

places, garage spaces, stories and the total square feet of the house. Additionally, the

paper adds three dummy variables, MAY, JUNE, and JULY, to account for seasonable fac-

tor with regards to the houses’ prices. If the house is sold in May, the variable MAY is

set to be equal to 1 and 0 otherwise. The other variables, JUNE and JULY are constructed

in a similar fashion. The paper finds that the log sale prices of houses are significantly

and positively correlated with MAY and JULY while JUNE is insignificant. This implies

that houses that are closed in May or July tends to have a higher price. Similar to Cebula

(2009), my paper includes sold month of the house as dummy variables. However, these

attributes do not appear to be statistically significant.

3

Page 7: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

Selim (2009) seeks to study the effects of different housing characteristics on housing

prices in Turkey using two different methods: hedonic pricing model and artificial neural

network [15]. The paper’s dataset, which was collected from the 2004 Household Budget

Survey Data for Turkey, contains 5,741 observations with 46 housing characteristics. For

the hedonic pricing model, the author uses the semi-log form, ln(P ) = βx + u, where

P denotes the price of the house, x is the set of independent variables and u is the error

term. As for the artificial neural network model, the paper uses 2 hidden layers, with nine

and four nodes for the first and second layer, respectively. The results are consistent with

other studies on housing price. The author finds that the total number of rooms, the size

of the house, the heating systems, appliances such as garbage disposal, garage and pool,

etc. have a significant and positive effect on the house price. More importantly, Selim

finds that the artificial neural network model has a lower error score than the hedonic

model. When the hedonic model’s mean squared error is 2.47, the same error measure-

ment by the neural network model is 0.44. Similarly, Tay and Ho (1991/1992) compared

the pricing prediction between regression analysis and artificial neural network in pre-

dicting apartments’ prices in Singapore [18]. They found that the neural network model

outperforms regression analysis model with a mean absolute error of 3.9%.

Jirong, Mingcang, and Liuguangyan (2010) uses support vector machine (SVM) regres-

sion to forecast the housing prices in China in between 1993 and 2002 and in certain dis-

trict in Tangshan city in between 2000 to 2002 [9]. The paper utilizes the genetic algorithm

to tune the hyper-parameters in the SVM regression model. The error scores for the SVM

regression model for both China and a Tangshan City’s district are both lower than 4%.

This indicates that the SVM regression model perform well in forecasting housing prices

in China. In the Singapore’s housing market, Fan, Ong and Koh (2006) uses decision tree

model study the housing characteristics’ effects on prices [6]. The paper concludes that

the owners of 2-room to 4-room flats are more concerned with the flats’ basic characteris-

4

Page 8: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

tics such as model type and age more than the owners of 5-or-more-room flats. Moreover,

owners of executive flats care more about the services characteristics such as the neigh-

borhood location and recreational facilities than basic housing characteristics.

3 Methods

3.1 Hedonic Pricing Model

In Economics, the hedonic pricing model is frequently used to measure a property’s price.

The model is based on the theory of consumer’s demand by Lancaster (1966), which states

that utilities of a good is not based on the good itself but on the individual “characteris-

tics” of the good [10]. However, it’s not until Rosen (1974) that the idea of pricing is added

to the model. Rosen argues that a good can be evaluated based on the individual values

of its composite attributes [14]. Since then, this pricing model has been adapted to evalu-

ate properties based on their internal and external characteristics.

Hedonic pricing model combines both a house’s internal characteristics (such as number

of bedrooms, number of bathrooms, etc.) and its external characteristic (such as neighborhood’s

walkability score, public schools’ scores, etc.) to estimate its values. A hedonic model can be

written as a linear regression model, as follows:

Pi =k∑

i=1

wi,mEi,m +k∑

i=1

wi,nIi,n + b (1)

In equation (1), there are k observations with m number of External housing attributes

and n number of Internal attributes. Moreover, b represents the constant term. This model

explores the linear relationship between various characteristics of a house and its actual

sold price. For example, if the coefficient of the variable “bathroom” (w) in the hedonic

model is 15000, keeping other variables constant, if a house has one more bathroom, its

5

Page 9: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

sold price could go up by $15,000.

In this project, the hedonic pricing model or Linear Regression is used as the baseline

model to compare more complex machine learning algorithms against. This model is

chosen for its frequent appearance in Economics papers on housing price prediction and

its simplicity in explaining relationships among attributes.

3.2 Machine Learning Algorithms

This project uses WEKA2, a suite of machine learning algorithms, developed at the Uni-

versity of Waikato. There are various algorithms3 tested, based on their abilities to handle

regression analysis and their appearances in previous literature. The best performing

ones are Random Forest and Support Vector Regression, which are explained in details in

the next two subsections.

3.2.1 Random Forest

Random Forest is a learning algorithm first created by Tin Kam Ho [7], a computer sci-

entist at IBM, and later extended by Leo Breiman and Adele Cutler [2] [13]. It operates

by constructing a multitude of decision trees to fit the observations into groups based

on their attributes’ values and outputs the mean prediction of the individual trees. As

the name suggests, “decision tree” model builds a reversed tree-like structure, where the

“root” is at the top, followed by multiple branches, nodes and leaves. The end of each

branch is a decision leaf, which is the model’s predicted value, given the values of the

attributes represented by the path from the root node to the said decision leaf. Figure 1

presents a sample decision tree where the dependent variable or the decision leaf is the

sold price of a house, and the dependent variables or the nodes are the number of bath-

2https://www.cs.waikato.ac.nz/ml/weka/3These include IBk, Artificial Neural Network (Multilayer Perceptron), and Decision Tree.

6

Page 10: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

rooms, bedrooms, and the size of the house. In Figure 1, the leftmost decision leaf (Price =

$50,000) is derived using the path in which the root node “Baths” has a value of less than

two and the node “Beds” has a value of less than or equal to two. This tree’s maximum

depth, which can be defined by the longest distance from a decision leaf to the root node,

is therefore three. In building a decision tree, the best attribute of the dataset, in terms of

error deduction, is placed at the top of the tree (root node). The next node used in each

tree branch produced by the root node is considered the next best attribute if the attribute

at the root is removed from the dataset, and so forth. The process of choosing which node

to use at each tree branch is described below:

1. For each of the independent variables, fit a regression between the independent and

the dependent variable.

2. For each of the independent variables, the observation set is split into several dis-

joint subsets at certain values of the variable.

3. At each split point, the error between the predicted and actual value is squared to

create the sum of squared errors (SSE).

4. The SSE is compared across all independent variables and the split points. The

variable/split point with the lowest SSE is chosen to be the root node and the split

point for the root node.

Figure 1: A sample decision tree model

7

Page 11: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

After following the four steps above, the tree in Figure 1 chooses the root node as “Baths”

with a split point of 2. This means that this variable along with the split point produces

the smallest SSE compared to other variables and split points. After identifying the root

node, the algorithm uses steps 1 to 4 again for each branch of the tree (when Bathrooms <

2 and when Bathrooms≥ 2) until all the data is processed and the decision leaves contain

house price. Based on the decision tree model from Figure 1, a house that has 1 bathroom

with 3 beds is estimated at $100,000 whereas a house with 3 bathrooms, size of 2,000

square feet and 3 bedrooms is estimated at $70,000.

For a dataset with many attributes, using decision tree can lead to a large number of splits,

which creates a large and complex tree. When a tree is designed so that it can fit all the

training data points too well, the over-fitting problem occurs. This leads to inaccuracy

when predicting value of data points in the testing sets. By using random subsets of

attributes or observations for training on different trees, Random Forest can therefore

limit the over-fitting problem. In addition, Random Forest works well with datasets with

missing values, using the “surrogate split” method4.

3.2.2 Support Vector Regression

Support Vector Machine (SVM) is developed by Vladimir N. Vapnik and Alexey Ya. Cher-

vonenkis, two Russian statistician/mathematician in 1963 [21]. In 1996, a version of the

algorithm, called Support Vector Regression (SVR), was introduced by Vladimir N. Vap-

nik, Harris Drucker, Christopher J. C. Burges, Linda Kaufman and Alexander J. Smola [5].

Instead of fitting a best fitted line over the observations like Linear Regression, SVR with

a particular kernel (a polynomial kernel with a degree of one), called Linear SVM Regres-

sion, fits a flat hyperplane. This Linear SVM is used for the project. Figure 2 shows an

example of fitting a hyperplane through a collection of data points in three dimensions.

4Subsection 7.1 under the Appendix Section provides an explanation of how surrogate split works.

8

Page 12: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

For any point within the margins of the hyperplane, its error would be 0. The model is

described as yi = wxi+ b, in which yi is the predicted value, xi is the attribute and w is the

weight of the attribute xi.

Figure 2: Hyperplane Fitting Figure 3: Linear Support Vector Regression

The goal of SVR is to minimize 12||w||2, which helps reduce over-fitting in training data

and therefore reduces test errors [17]. The error is calculated as shown in equation (2).

Error =N∑i=1

[max(yi − (wxi + b)− ε, 0)] (2)

In equation (2), ε is the margin of the hyperplane as shown in Figure 3, and yi, xi and w

are as explained above.

By minimizing the weights of attributes, Support Vector Regression is less sensitive to

noises. Therefore, this model works well in finding patterns in real-life noisy datasets,

such as financial data.

9

Page 13: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

4 Data

4.1 Data Collection

This project collects data on 1,457 sold houses from Zillow, Trulia and Redfin using Python

and Selenium (a browser automation tool) for data scraping. These houses are selected

from five counties in five different regions of the United States. Figure 4 shows the ge-

ographical locations of these counties on the United States’ map along with their corre-

sponding states, regions and the number of houses scraped from them.

Figure 4: Selected Counties On The US Map

These five counties are chosen based on the following criteria:

1. They are from different regions in the United States.

2. They are among the worst performing counties based on the percentage of houses

whose Zestimates fall within the 5% range of their actual sold prices. This suggests

that there could be visible improvements in the price prediction algorithms for these

particular counties. The list of how well Zestimates perform in different counties

10

Page 14: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

can be found on Zillow’s website.

3. These counties’ housing information is available on all three housing websites.

Since the result for this project is calculated as the percentage of houses whose predicted

prices fall within the 5% range of their actual sold prices, the corresponding evaluation

baseline is the percentage of houses whose Zestimates are within the 5% range of their

actual sold prices. Therefore, it is important to scrape the Zestimate prediction for every

sold house. However, Zillow updates the Zestimate number even after a house is sold,

using the Zestimate currently listed on Zillow’s page would be inaccurate. Instead, the

historic Zestimate prediction right before a house is sold is needed. Since Zillow keeps

track of the Zestimate history for every month, with houses’ Zillow property ID, I was

able to find the Zestimate number for the month right before a particular house is sold,

and use this to compute my data’s prediction score baseline.

County State Region # of HousesZillow’s Countywide

BaselineMy Data’sBaseline

Cayuga NY Northeast 399 16.7 28.6Montgomery IL Midwest 209 8.7 21.1Upson GA Southeast 310 10.7 13.9Hunt TX Southwest 195 19.8 39.5Cowlitz WA West 354 29.3 27.7

Table 1: Selected Counties’ Information

Table 1 summarizes the five counties’ information, including the Zillow’s Countywide Base-

line for each county, which is the percentage of houses on Zillow whose Zestimates fall

within 5% of the houses’ actual sold prices. This is also the number Zillow reported on

their website as a measurement of their algorithms’ price predictability on each county.

Since my dataset only contains a fraction of houses that Zillow has, I compute my own

data’s baseline, which is shown in Table 1 as My Data’s Baseline. My Data’s Baseline is

11

Page 15: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

measured in a similar way as Zillow’s Countywide Baseline, only on the dataset of 1,457

houses.

4.2 Data Processing

This project collects 1,457 houses from three prominent different housing websites (Zil-

low, Trulia, and Redfin). The reason is to make sure that the scraped housing data is as

accurate as possible. Since all three websites get data from listing services or other third-

party companies, inconsistency and mistakes in data are unavoidable. Figure 5, which

shows how the same house (215 P G Street Rd, Kelso, WA 98626) is recorded across Zillow,

Trulia and Redfin, will serve as an example of inconsistency in sold price.

Figure 5: An Example Of Data Inconsistency

All three websites record that the house was last sold in October, 2017. However, the

sold price of the houses is inconsistent. On Zillow and Redfin, the house’s sold price is

recorded as $325,000 while the price is $233,427 on Trulia. In order to handle this consis-

tency, I simply pick the number that appears most often (which, in this case, would be

$325,000). If the three websites record three different sold prices, then I take an average

12

Page 16: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

of the three and mark the house for a later check-up. In case that the sold prices vary in

a wide range across the three websites (e.g. $1 in Zillow, $16,000 in Trulia and $58,000 in

Redfin), the observation will be drop. Besides sold price, other housing characteristics are

also subject to the same comparison algorithm.

Besides inconsistency in data values across three websites, there is also inconsistency in

data units. For example, size of a house can be recorded in either square feet or acres.

Therefore, an extra conversion step has to be taken in order to uniform data units. All

data processing steps are done in Excel’s Visual Basic for Applications (VBA).

4.3 Data Description

In this dataset, there are 35 housing attributes, including internal attributes and external

attributes. Internal housing attributes, such as number of bedrooms and number of bathrooms,

are intrinsic variables to the houses. On the other hand, external housing attributes, such

as the walkability of the neighborhood and public schools’ scores, are variables that are not built-

in with the houses. Table 6 and 7 in the Appendix Section show a full list of attributes,

both numeric and non-numeric, used in this project. If the attributes are numeric, Table

6 displays their summary statistics. If the attributes are non-numeric, Table 7 lists these

attributes and the coded dummy/binary variables these attributes are turned into. For

example, for the non-numeric attribute Sold Month, its dummy variables are the twelve

months of the year. If a house is sold in January, then the variable January would take a

value of 1, and 0 otherwise.

In this project’s dataset, one of the attributes is comparable houses’ sold price. In this pa-

per, this attribute is recorded as Comparables’ Sold Price or Coms’ Sold Price, for short. Zil-

low, Trulia and Redfin all provide a list of comparable houses (based on similar features

such as location, square footage and beds/baths) to the house currently being looked at,

13

Page 17: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

called “House X”. If a comparable house is sold before “House X” and the sold date is

within one year of the sold date of “House X”, the price of this comparable house is put

in a list of comparable houses’ sold price. For every house, there are three lists of the

comparable houses’ sold price from the three housing websites. The attribute Compara-

bles’ Sold Price is then calculated as the average of these three lists’ median values.

5 Results

5.1 Prediction Scores

Figure 6 shows the prediction scores of the three algorithms used in this project (Linear

Regression (LR), Support Vector Regression (SVR) and Random Forest (RF)) in compar-

ison to the data’s baselines5 for all five counties. The county dataset with asterisk (*)

next to it demonstrates that the best performing algorithm is statistically better than the

baseline algorithm, or LR. As mentioned before, the prediction score is measured as the

percentage of houses whose estimated prices fall within the 5% range of their actual sold

prices.

On the dataset of Hunt (TX), SVR outperforms the baseline by 3.2%. Zestimate predicts

39.5% of houses in Hunt county’s data to be within the 5% range of their actual sold

prices while SVR estimates 42.7% of houses to be in this 5% range. Moreover, the result

produced by SVR is statistically better than LR, the baseline algorithm. For the dataset of

Upson (GA), RF gives the same prediction score as the baseline (13.9%). In addition, RF

produces prediction scores that are close to the baselines for the datasets of Cowlitz (WA)

and Montgomery (IL). For these two counties, the differences between RF’s predictions

and the baselines are around 3%. However, RF is not statistically different than LR for

the datasets of the three counties in Washington, Illinois, and Georgia. Among the five

5This data’s baselines are the same as the values recorded in the last column of Table 1.

14

Page 18: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

Figure 6: Models’ Prediction Scores Compared To Baselines

counties, Cayuga (NY) appears to have the worst performance. For this county’s dataset,

SVR is the best performing algorithm, followed by RF and LR. For Cayuga (NY), the pre-

diction score gap between the baseline and SVR’s performance is 11%. However, SVR is

statistically better than LR for this particular county.

In order to produce the results as shown in Figure 6, each county data has a different

set of attributes that are considered “most important” in terms of predicting sold prices.

These attributes are selected using a combination of WEKA’s Attribute Selected Classifier,

which evaluates the predictability of a subset of attributes by considering the individual

predictability of each attribute and the degree of redundancy between them, and through

“trial and error” experiments. Table 8 in the Appendix Section shows these most impor-

tant attributes for each of the five counties. However, it would be beneficial to have the

same set of attributes for all counties, so that we can have a uniform frame of reference

for comparison of attributes’ effects on sold prices. Given the sets of most important at-

15

Page 19: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

Figure 7: SVR - Different Attributes vs. Similar Attributes

Figure 8: RF - Different Attributes vs. Similar Attributes

16

Page 20: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

tributes across counties, I picked out 10 attributes that are important to the majority of the

counties. In another words, these 10 attributes appear at least 3 times within the 5 coun-

ties’ datasets. For example, assessment appears as the one of the most important attributes

in all five counties’ datasets and number of bedrooms appears in three datasets. Therefore,

these two attributes are included in the list of 10 most important attributes across the five

counties. Table 9 in the Appendix Section shows a list of these attributes and which coun-

ties’ datasets they appear on.

Figure 7 and 8 display the prediction score comparison between having one common set

of attributes for 5 counties and having different sets of attributes for different counties.

Figure 7 uses SVR as the algorithm whereas Figure 8 uses RF. These figures suggest that

switching from different set of attributes to a single set don’t change the prediction scores

by a lot. In some cases, such as the dataset of Hunt (TX) with SVR (Figure 7), using the

same set of attributes yields a better result.

5.2 Overestimation Problem

As mentioned before, Zillow tends to overestimate their properties. The Introduction

Section states the reason why over-pricing houses can have a negative effect on the final

sale prices. In this dataset of 1,457 houses, the ratio of overestimated to underestimated

houses by Zillow is 3 to 2 (Figure 9a). However, my models successfully reduce this ratio

to 1 to 1. Figure 9b shows the ratio of overestimated to underestimated houses when using

SVR while Figure 9c shows the ratio when using RF as the algorithm. The horizontal

axis represents price difference range (in thousand dollars) while the vertical axis shows

the percentage of houses that fall within a certain price difference range. Negative price

difference, which means the Zestimate/predicted price is lower than the actual sold price,

is put in brackets, as shown on the horizontal axis of the graph.

17

Page 21: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

(a) Zestimate Gives a Ratio of 3:2

(b) Support Vector Regression Gives a Ratio of 1:1

(c) Random Forest Gives a Ratio of 1:1

Figure 9: Overestimated To Underestimated House Ratio Comparison

18

Page 22: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

5.3 Attributes’ Effects On Sold Price

Table 2 shows the coefficients of the most important attributes across 5 counties that are

also statistically significant at 95% confidence level when it comes to predicting house

prices. The coefficients shown in the table are calculated by taking average of the at-

tributes’ coefficients among the counties. Results show that, on average, $100 increase in

the house’s assessment could increase its sold price by $54. The same amount increase

in the sold price of comparable houses and the house’s listed price would increase the

sold price by $34 and $38, respectively. Moreover, Table 2 suggests that if a house has one

more bathroom, its sold price could go up by $15,787.

ID Attribute Coefficient1 Assessment 0.542 Comparables’ Sold Price 0.343 Listed Price 0.384 Baths 15,787

Table 2: Most Important and Statistical Significant Attributes

The effects of these four attributes can vary across counties. For example, the coefficient

of Baths is higher in Upson (GA) than in Montgomery (IL), suggesting that the number

of bathroom has a bigger effect on sold price in Upson county. Table 3 shows the effects

of these four attributes on houses’ sold prices for each of the five counties. A blank entry

in the table means that the attribute is not considered an “important” attribute for that

particular county.

Attribute1 Cayuga (NY) Cowlitz (WA) Hunt (TX) Montgomery (IL) Upson (GA)Assessment 0.39 0.39 0.15 0.57 1.22Coms’ Sold Price 0.30 0.51 0.12 0.47 0.29Listed Price 0.27 0.23 0.62 0.40Baths 16,697 12,650 18,014

1 These attributes are statistically significant at the 95% confidence level.

Table 3: Attributes’ Effects On Sold Price Across Selected Counties

19

Page 23: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

6 Conclusion

Using a dataset of 1,457 houses from 5 different counties scraped from Zillow, Trulia and

Redfin, this paper addresses the following questions:

1. Can the models proposed in this paper outperform or get close to Zillow’s predic-

tion score baseline?

2. Can the overestimated to underestimated house ratio be reduced?

3. What are the most important attributes that affect sold price?

For Hunt (TX), SVR outperforms the baseline by 3.2%. RF outputs close predictions scores

to the baseline with the dataset from Cowlitz (WA) and Montgomery (IL). The differences

between RF’s predictability and Zestimate for these two counties is around 3%. RF gives

a similar score as the baseline for Upson (GA). Moreover, results suggest that using one

single set of 10 attributes for all counties will not change the models’ accuracy scores by a

lot in comparison to using different sets of attributes for different counties. The overesti-

mated to underestimated house ratio is also reduced from 3:2 to 1:1. In addition, the four

most important and statistical significant attributes are identified as number of bathrooms,

assessment, listed price and comparable houses’ sold price.

Finally, for future work, it would be interesting to see what results could be yielded from

applying the same models on counties that Zillow reports to be the best performing ones.

20

Page 24: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

7 Appendix

7.1 Surrogate Split

Consider a dataset of 10 observations, 3 independent variables (Baths, Beds, and Rooms)

and the dependent variable as Sold Price, as shown in Table 4 below. Assume that when

we apply a tree algorithm to this dataset, Baths is picked as the tree’s root node with a

split point of 1. This is called the “primary split”. However, the algorithm still have to

determine whether to put x5 and x8 in the group of observations with fewer than or equal

to 1 bath (x2, x6) or in the group with more than 1 bath (x1, x3, x4, x7, x9, x10), because Baths

is missing for both of these observations.

Obs. Baths Beds Roomsx1 2 4 1x2 1 ? 3x3 3 3 4x4 2 3 2x5 ? 3 4x6 1 2 ?x7 3 3 5x8 ? 3 4x9 2 3 3x10 2 1 4

Table 4: Sample Data

The surrogate splitting method is then utilized. It uses the attributes Beds and Rooms

to create “alternative splits”. If the best split point using the attribute Beds is 3, we

will have two groups of observations: (x6, x10) and (x1, x3, x5, x7, x8, x9, x4). If the best

split point using the attribute Rooms is 4, we have two different sets: (x1, x2, x4, x9) and

(x3, x5, x7, x8, x10). Comparing between the sets produced by using the attributes Rooms

and Beds, we see that the two sets produced by Beds bear more resemblance to the two sets

produced by the attribute Baths in terms of the number of observations that are grouped21

Page 25: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

together. Therefore, we say that Beds produces the best “surrogate split”. Using this “sur-

rogate split”, the two observations x5 and x8 are grouped with observations x1, x3, x7, and

x9, which, in the “primary split”, are pooled together in the group that has more than 1

bathroom. Therefore, after using the surrogate splitting method, the tree algorithm knows

which tree branch to put observations with missing values into.

7.2 Tables

• Table 5 shows a summary of the literature, specifying the previous works’ datasets,

methods, accuracy and error scores.

• Table 6 shows a list of the numeric attributes.

• Table 7 shows a list of the non-numeric attributes.

• Table 8 shows the most important attributes for each county.

• Table 9 shows the ten most important attributes across all five counties.

22

Page 26: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

Tabl

e5:

Lite

ratu

reSu

mm

ary

Pape

rD

ata

Met

hod

Acc

urac

yEr

ror

Ceb

ula

(200

9)2,

888

sing

lefa

mily

hous

esin

Geo

rgia

Hed

onic

Pric

ing

Mod

elR

-squ

ared

=0.

86

Fan,

Ong

,and

Koh

(200

6)5,

589

flats

inSi

nga-

pore

+D

ecis

ion

Tree

.R

-squ

ared

=0.

88

+Tr

aini

ng-T

estp

er-

cent

age

is75

%-2

5%.

+U

ses

Tree

Nod

ein

SAS,

whi

chin

tegr

ates

the

CH

AID

,CA

RT,

and

C4.

5/C

5.0

algo

-ri

thm

s.Ji

rong

,Min

gcan

g,an

dLi

ngua

ngya

n(2

011)

+N

atio

nala

nnua

lse

lling

pric

ein

Chi

nabe

twee

n19

93an

d20

00.

+Su

ppor

tVec

tor

Ma-

chin

e.+

Mea

nA

bsol

ute

Per-

cent

age

Erro

r:2.

33%

for

Chi

nada

ta

+Q

uart

erly

selli

ngpr

ice

ofTa

ngsh

anci

tybe

twee

n20

00an

d20

02

+U

ses

Gen

etic

Al-

gori

thm

sto

opti

miz

eSV

M.

+M

ean

Abs

olut

ePe

r-ce

ntag

eEr

ror:

1.94

%fo

rTa

ngsh

anC

ity

data

Tay

and

Ho

(199

1/19

92)

1,05

5ob

serv

atio

ns+

AN

N(1

hidd

enla

yer

wit

h5

node

s)+

Mea

nA

vera

geEr

-ro

rfo

rA

NN

=3.

9%+

Reg

ress

ion

+M

ean

Ave

rage

Er-

ror

for

Reg

ress

ion

=7.

5%Se

lim(2

009)

5,74

1ho

uses

inTu

rkey

+A

NN

(2hi

dden

laye

rs,w

ith

9an

d4

node

sea

ch)

+M

ean

Abs

olut

eEr

-ro

rfo

rA

NN

=0.

51

+H

edon

icPr

icin

gM

etho

dH

edon

icPr

icin

gM

odel

’sR

-squ

ared

=0.

65

+M

ean

Abs

olut

eEr

-ro

rfo

rH

edon

icPr

ic-

ing

Mod

el=

1.23

23

Page 27: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

ID Attribute Percentage Missing Min Max Mean Std. Dev.1 Sold Price 0 1,127 695,000 141,380 105,3722 Sold Year 0 2015 2018 1026.9 0.43 Beds 0 1 7 3.0 0.84 Baths 0 1 6 1.8 0.75 Size (sqft) 0 458.5 7481 1715.6 751.96 Lot (sqft) 9 1742 958,320 65,696 125,9757 Date Built 6 1790 2017 1957 40.58 Last Remodel Year 65 1900 2017 1977.5 19.19 Tax Amount 9 65 12,246 1,996 1,580

10 Assessment 7 850 545,420 104,722 88,49111 Elementary School Score 1 0 9 4.4 1.712 Middle School Score 0 0 9 4.9 1.613 High School Score 0 0 9 4.7 1.414 Previous Sold Price 1 66 1 510,000 118,452 84,59915 Listed Price 41 10,400 850,000 179,605 119,88616 Changed Price 76 10,400 724,999 174,386 129,46317 Walkability 11 0 82 20.7 21.718 Rooms 71 1 17 6.7 2.619 Garage Space 62 1 17 2.1 1.320 Floor 51 1 5 1.3 0.521 Coms’ Sold Price 21 2,000 513,276 142,294 90,58322 Restaurants 18 0 76 11 1423 Grocery Stores 18 0 13 1 224 Nightlife 18 0 21 2.5 4.6

1 There are two houses in Cayuga, NY whose previous sold prices are $1. These are considered anomalies but thetwo houses are still kept in the dataset because their current sold prices and other attributes are within normalranges.

Table 6: Numeric Attributes

24

Page 28: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

ID Attribute Representatives1 Sold Month Jan, Feb, etc.2 House Type Single Family, Condo, Townhouse3 Street Parking Yes, No4 Floor Type Hardwood, Carpet, Tile5 Heating System Gas, Electric, Center6 Cooling System Center, Electric7 Appliances Dishwasher, Garbage Disposal, Oven, Fridge8 External Material Stone, Wood, Brick9 House Style Colonial, Contemporary, Ranch, Bungalow10 Roof Type Asphalt, Composition, Metal, Shake Single11 Garage Type Detached, Attached

Table 7: Non-Numeric Attributes

25

Page 29: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

Tabl

e8:

Mos

tIm

port

antA

ttri

bute

sFo

rSe

lect

edC

ount

ies

IDC

ayug

a(N

Y)

Cow

litz

(WA

)H

unt(

TX

)M

ontg

omer

y(I

L)U

pson

(GA

)1

Size

Size

Size

Beds

Beds

2D

ate

Built

Tax

Am

ount

Lot

Bath

sBa

ths

3La

stR

emod

elYe

arA

sses

smen

tD

ate

Built

Size

Lot

4Ta

xA

mou

ntLi

sted

Pric

eTa

xA

mou

ntD

ate

Built

Dat

eBu

ilt5

Ass

essm

ent

Elec

tric

Hea

ting

Ass

essm

ent

Tax

Am

ount

Last

Rem

odel

Year

6Li

sted

Pric

eC

ente

rC

oolin

gLi

sted

Pric

eA

sses

smen

tA

sses

smen

t7

Wal

kabi

lity

Asp

halt

Roo

fW

alka

bilit

yPr

evio

usSo

ldPr

ice

Com

posi

tion

Roo

f8

Dis

hwas

her

Com

s’So

ldPr

ice

Har

dwoo

dFl

oor

List

edPr

ice

Com

s’So

ldPr

ice

9G

arba

geD

ispo

sal

Beds

Tile

Floo

rW

alka

bilit

yC

hang

edPr

ice

10C

onte

mpo

rary

Styl

eBa

ths

Floo

rD

ishw

ashe

rW

alka

bilit

y11

Com

s’So

ldPr

ice

Elem

enta

rySc

hool

Scor

eC

oms’

Sold

Pric

eR

esta

uran

tsD

ishw

ashe

r12

Prev

ious

Sold

Pric

eM

iddl

eSc

hool

Scor

eBu

ngal

owSt

yle

Ove

n13

Hig

hSc

hool

Scor

eC

oms’

Sold

Pric

eR

esta

uran

ts14

Gro

cery

Stor

es15

Nig

htlif

e

26

Page 30: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

ID Attribute Cayuga (NY) Cowlitz (WA) Hunt (TX) Montgomery (IL) Upson (GA)1 Assessment x x x x x2 Coms’ x x x x x

Sold Price3 Date Built x x x x4 Listed Price x x x x5 Size x x x x6 Tax Amount x x x x7 Walkability x x x x8 Baths x x x9 Beds x x x

10 Dishwasher x x x

Table 9: Set Of Most Important Attributes

27

Page 31: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

References

[1] Paul K. Asabere and Forrest E. Huffman. “Price Concessions, Time of the Market,

and the Actual Sale Price of Homes”. In: Journal of Real Estate Finance and Economics 6

(1993), pp. 167–174. URL: https://link.springer.com/article/10.1007/

BF01097024.

[2] Leo Breiman. “Random forests”. In: Machine learning 45.1 (2001), pp. 5–32.

[3] Rochard J. Cebula. “The Hedonic Pricing Model Applied to the Housing Market of

the City of Savannah and Its Savannah Historic Landmark District”. In: The Review

of Regional Studies 39.1 (2009), pp. 9–22. URL: journal.srsa.org/ojs/index.

php/rrs/article/download/182/137.

[4] Consumer Housing Trends Report 2016. Zillow Group. Accessed: 11/10/2017. 2016.

URL: https://www.zillow.com/research/zillow- group- report-

2016-13279/.

[5] Harris Drucker et al. “Support vector regression machines”. In: Advances in neural

information processing systems. 1997, pp. 155–161.

[6] Gang-Zhi Fan, Seow Eng Ong, and Hian Chye Koh. “Determinants of House Price:

A Decision Tree Approach”. In: Urban Studies 43.12 (2006), pp. 2301–2315. URL:

journals.sagepub.com/doi/pdf/10.1080/00420980600990928.

[7] Tin Kam Ho. “Random decision forests”. In: Document analysis and recognition, 1995.,

proceedings of the third international conference on. Vol. 1. IEEE. 1995, pp. 278–282.

[8] Daniel R. Hollas, Ronald C. Rutherford, and Thomas A. Thomson. “Zillow’s es-

timates of single-family housing values.” In: Expert Systems with Applications 78.1

(2010). URL: http://www.freepatentsonline.com/article/Appraisal-

Journal/220765044.html.

28

Page 32: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

[9] Gu Jirong, Zhu Mingcang, and Jiang Liuguangyan. “Housing price based on ge-

netic algorithm and support vector machine”. In: Expert Systems with Applications

38 (2011), pp. 3383–3386. URL: http://www.sciencedirect.com/science/

article/pii/S0957417410009310.

[10] Kelvin J. Lancaster. “A New Approach to Consumer Theory”. In: The Journal of Po-

litical Economy 74.2 (1966), pp. 132–157. ISSN: 0303-2647. DOI: 10.1.1.456.4367&

rep=rep1&type=pdf. URL: http://www.jstor.org/stable/1828835.

[11] Number of houses sold in the United States from 1995 to 2016. www.statista.com. Ac-

cessed: 11/10/2017. URL: https://www.statista.com/statistics/219963/

number-of-us-house-sales/.

[12] Quick Facts: Resident Demographics. National Multifamily Housing Council. Accessed:

11/11/2017. 2017. URL: http://www.nmhc.org/Content.aspx?id=4708.

[13] Random Forests by Leo Breiman and Adele Cutler. URL: https://www.stat.berkeley.

edu/˜breiman/RandomForests/.

[14] Sherwin Rosen. “Hedonic Prices and Implicit Markets: Product Differentiation in

Pure Competition”. In: The Journal of Political Economy 82.1 (1974), pp. 34–55. URL:

http://people.tamu.edu/˜ganli/publicecon/rosen74.pdf.

[15] Hasan Selim. “Determinants of house prices in Turkey: Hedonic regression versus

artificial neural network”. In: Expert Systems with Applications 36 (2009), pp. 2843–

2852. URL: www.sciencedirect.com/science/article/pii/S0957417408000596.

[16] G. Stacy Sirmans, David A. Macpherson, and Emily N. Zietz. “The Composition of

Hedonic Pricing Models”. In: Journal of Real Estate Literature 13.1 (2005), pp. 3–43.

URL: http://www.jstor.org/stable/44103506?seq=1#page_scan_tab_

contents.

[17] Alex J Smola and Bernhard Scholkopf. “A tutorial on support vector regression”.

In: Statistics and computing 14.3 (2004), pp. 199–222.29

Page 33: Housing Price Prediction - Semantic Scholar · Housing Price Prediction An Nguyen March 20, 2018 ... neighborhood). Using data from sold houses listed on Zillow, Trulia and Redfin,

[18] Danny P. H. Tay and David K. H. Ho. “Artificial Intelligence and the Mass Ap-

praisal of Residential Apartments”. In: Journal of Property Valuation and Investment

10.2 (1992), pp. 525–540. URL: http://www.emeraldinsight.com/doi/abs/

10.1108/14635789210031181.

[19] The Price of Overpricing: How Listing Price Impacts Time on Market. Zillow. Accessed:

03/06/2018. 2016. URL: https://www.zillow.com/research/overpricing-

impacts-time-market-12476/.

[20] Trulia Estimate. Trulia. Accessed: 11/11/2017. 2017. URL: https://www.trulia.

com/trulia_estimates/.

[21] Vladimir N. Vapnik and Alexey Ya Chervonenkis. “On a class of algorithms of

learning pattern recognition.” In: Automation and Remote Control 25.6 (1964).

[22] Zestimate. Zillow Group. Accessed: 11/11/2017. 2017. URL: https://www.zillow.

com/zestimate/.

30


Recommended